Installation

Tangent is distributed on PyPI as tangent-ad. The import name is unchanged — your code says import tangent, exactly like the original Google release (the same arrangement as Pillow, which installs as pillow and imports as PIL).

pip install tangent-ad

Upgrading from Google's 2017 tangent package

PyPI's tangent project is Google's unmaintained 0.1.9 release from 2017. Both packages install a tangent/ module, so if you have the old one, remove it first:

pip uninstall tangent
pip install tangent-ad

Optional backends

The core install differentiates NumPy code. Each additional backend is an extra:

pip install "tangent-ad[jax]"       # JAX
pip install "tangent-ad[tf]"        # TensorFlow 2.x
pip install "tangent-ad[torch]"     # PyTorch
pip install "tangent-ad[keras]"     # Keras 3 (backend-agnostic keras.ops)
pip install "tangent-ad[tinygrad]"  # tinygrad
pip install "tangent-ad[viz]"       # gradient-flow visualization (matplotlib + networkx)
pip install "tangent-ad[symbolic]"  # SymPy-based algebraic simplification
pip install "tangent-ad[all]"       # everything above, plus the test tooling

Backends are optional by design: extensions load when the backend is importable and stay silent otherwise. Core NumPy autodiff always works.

From source

pip install "tangent-ad @ git+https://github.com/pedronahum/tangent.git"

or for development:

git clone https://github.com/pedronahum/tangent.git
cd tangent
pip install -e ".[test]"
python -m pytest tests/ -q --short

Requirements

• Python 3.9 – 3.13 (tested in CI on all five)
• numpy and gast (installed automatically)

Platform note: TensorFlow has no CUDA build for aarch64; on such machines the TF backend runs on CPU.